OpenRouter

Private Endpoints lets an organization register a fine-tuned model, dedicated deployment, or reserved capacity as a private endpoint and call it through the existing OpenRouter API and model slug, with public providers usable as fallback and the endpoint hidden from public model lists and rankings. The feature is an Enterprise add-on available through account representatives, with a settings-page wizard and a Management API key-authenticated Private Endpoints API covering registration, testing, activation, pricing, pausing, and deletion, wrapped by an openrouter_private_endpoint Terraform resource.

OpenRouter introduced a Model Router Benchmarks page that scores routers across varied domains and synthesizes quality, speed, and cost into a 0 to 10 Router Index, with a slider to adjust the default 60/20/20 weighting. The page lists benchmarks for routers including Auto, Fugu Max, Fugu Ultra v2, Jev Router, Pareto, and NVIDIA Switchyard, alongside individual models as baselines, and excludes alias-style routers like Pareto Code, -latest slugs, and Fusion.

OpenRouter's new Security Center shows every key across workspaces in one place, scoring risk by spend limit and expiration, with bulk actions to disable, archive, or cap up to 500 keys at once and a CSV export of any recommendation's keys and owners. The IP allowlist, previously in privacy settings, gets its own tab and now applies to every key in an account or organization, rejecting off-list requests with a 403 Forbidden.

OpenRouter launched an asynchronous batch API at api/v1/batches supporting chat completions, responses, messages, and embeddings, with providers generally charging 50% or less of normal per-token pricing in exchange for completing requests within a 24-hour window. Across 230k+ beta batches, the median finished in 7 minutes and 90% within an hour.

Requests to us.openrouter.ai and eu.openrouter.ai are decrypted inside the region and routed only to providers operating there, and a request fails with a 404 rather than falling back outside the region when no in-region provider serves the model. In-Region Routing is available on the Business and Enterprise plans, and can be enforced per workspace, team, or API key with Guardrails.

OpenRouter launches the openrouter:shell server tool and a Files API, letting any model that supports tool calling run commands in a hosted Linux container, upload files to work with, and download outputs. Both are in beta, with sandbox time billed at $0.0001 per second as part of the request.

OpenRouter launched Visual Image Benchmarks showing every result from challenging prompts across 39 image models in a sortable grid by price and generation time, with the same side-by-side format extended to video and speech models. Prompts are grouped into seven families covering improbable scenes, counting, text rendering, spatial relations, negation, editing, and consistency.

OpenRouter's new Activity dashboard breaks down spend, requests, tokens, cache hit rate, and blended cost per million tokens across agents, models, API keys, and team members, with Overview, Trends, Explore, and Guardrails views that link through to filtered logs and a Generation detail view covering cost, latency, routing, and attribution. All data is available through the Analytics API and an openrouter-analytics skill for querying usage from an agent or terminal.

OpenRouter published live leaderboards across BrowseComp, DeepSearchQA, WideSearch, and HLE that rank model, engine, and search-budget combinations by quality, value, and speed. Findings show raising the search budget from 1 to 25 turns roughly doubles scores at 2.5-7x the cost, and that model choice moves results more than engine choice.

The Auto router for openrouter/auto now selects models using the past 7 days of aggregate spend on OpenRouter, classifying each prompt into one of roughly 30 task types and routing to the models the community spends on for that task. A cost_tier parameter (low, medium, high, xhigh, max) sets the spend band, and sticky routing keeps multi-turn conversations on one model while it remains a leading choice.

Classifiers, in beta, let workspaces define a taxonomy of up to eight dimensions and have a chosen model asynchronously tag each generation, or a sampled subset, with structured metadata written to logs. Six preset templates cover department, audience, task type, engineering work, agent complexity, and capitalizable software expense, and classifier dimensions can be grouped in the Activity Explorer.

Fusion dispatches a prompt to a panel of models in parallel with web search and web fetch enabled, then a judge model produces structured analysis that grounds the final answer, and it runs server-side behind a single openrouter/fusion model slug. On 100 DRACO deep research tasks, a Fable 5 + GPT-5.5 panel scored 69.0% versus Fable 5 solo at 65.3%, and a budget panel of Gemini 3 Flash, Kimi K2.6, and DeepSeek V4 Pro reached 64.7% at half the cost of Fable 5.

Product changes